Protein models, trained from scratch
NaviKinase 1.0 is a protein language model we pre-trained from scratch on 2.45 million kinase sequences. It is published openly so the capability can be checked rather than taken on trust.
NaviKinase 1.0
A decoder-only protein language model for the phosphotransferase superfamily — protein kinases plus sugar, lipid and nucleotide kinases. It was trained from scratch, not adapted from an existing model, and it speaks amino acid sequences rather than English.
It was trained entirely locally on Apple Silicon through MLX, with no cloud compute, running unattended over several days with checkpointing and crash recovery.
| Parameters | 149,604,416 (149.6M) |
|---|---|
| Architecture | Decoder-only transformer, pre-norm RMSNorm, RoPE, tied embeddings |
| Layers / width | 18 layers, d_model 832, 13 attention heads, FFN 3328 |
| Context | 1024 tokens |
| Training corpus | 2,450,504 sequences and 1,320,176,103 tokens from UniProtKB EC 2.7.-, deduplicated at the UniRef90 cluster level with cluster-level train/validation/test splits |
| Conditioning | Each sequence tagged with one of 14 EC 2.7.- sub-subclasses, so generation can be steered to a kinase family |
| Training run | Full 40,289-step ceiling (2 epochs) completed, no early stopping triggered |
| Held-out test loss | 1.4633 (perplexity 4.32) |
| Licence | CC BY-NC 4.0 — commercial use requires written permission |
| Published | huggingface.co/NavitraTechnologies01/navikinase-1.0 |
What it does
- Generates candidate kinase sequences from scratch, or continues a real one, optionally conditioned on an EC sub-subclass.
- Scores how well a real sequence fits what the model learned, per-position, to flag atypical stretches. This is a plausibility signal, not a classifier or a structure prediction.
- Scans for canonical catalytic motifs — the glycine-rich P-loop, catalytic HRD and activation-loop DFG — as a cheap directional check.
- Chains fetch, score, motif scan and literature search into a single research pass over a named protein, using live UniProt and PubMed lookups.
Why the test number is the one quoted
Validation loss was used for checkpoint selection, so it carries a small optimistic bias. The held-out test set was never touched during training or selection. That is the figure above, and it is the one to cite.
Responsible use. This model generates real biomolecules, not text. It has no content moderation, no toxin or hazard blocklist and no automated biosecurity screening, and it has not been reviewed by a biosecurity expert. Generated sequences are unvalidated hypotheses: anyone taking output toward wet-lab work is responsible for reviewing it independently and routing synthesis through a provider that performs standard biosecurity screening. The model’s scope — one of the most common and well-characterised enzyme families, with a closed EC-tagged vocabulary and no free-form prompting — limits the realistic risk, but that narrowness followed from data and model-size decisions rather than being an engineered safeguard, so it should not be treated as one.
Limitations
- Scope is EC 2.7.- only: no other transferase classes, non-enzymatic proteins or non-protein biomolecules.
- Single-sequence and character-level — no structural, alignment or homology information is used.
- Motif checks are a directional signal, not a validated functional or structural benchmark.
- The EC classification mode is an improvised proxy; the model was never trained as a classifier.
- Structure prediction is handled by ESMFold, a separate third-party model, and is not covered by the evaluation above.
What we can build for you
We build domain-specific sequence models for biotech and pharma teams, pre-trained from scratch on your own corpus using the same method as NaviKinase, and deployed inside your environment so proprietary sequences never leave it. See how NaviKinase compares with general protein models in our evaluation paper (PDF).
Commercial licensing of NaviKinase, or a model built on the same approach: get in touch.